On Thu, 24 Oct 2024 06:44:36 GMT, Emanuel Peter <[email protected]> wrote:
>> Jatin Bhateja has updated the pull request incrementally with one additional
>> commit since the last revision:
>>
>> Factor out IR tests and Transforms to follow-up PRs.
>
> test/jdk/jdk/incubator/vector/templates/Kernel-SaturatingBinary-Masked-op.template
> line 8:
>
>> 6:
>> 7: for (int ic = 0; ic < INVOC_COUNT; ic++) {
>> 8: for (int i = 0; i < a.length; i += SPECIES.length()) {
>
> I think this does not check if the generated vectors are too long. We had
> bugs in the past where we should have created say 2-element vectors, but the
> backend wrongly created 4-element vectors. This is especially an issue with
> vectors that do direct memory access.
>
> With a simple "counting-up" test, you will probably not catch this. It could
> be good to have a "counting-down" example as well. What do you think?
Also: all of these cases load, and directly store again. Does that not mean all
tests will probably pick the "..._mem" backend operations? Or do we actually
end up testing all backend operations with the tests we have here?
-------------
PR Review Comment: https://git.openjdk.org/jdk/pull/20507#discussion_r1814368049
